OCULARCENTRISM AND DEEPFAKES: SHOULD SEEING..., 31 Fordham Intell.... 31 Fordham Intell. Prop. Media & Ent. L.J. 1042 Fordham Intellectual Property, Media and Entertainment Law Journal Summer, 2021 Article Katrina Geddesa1 Copyright © 2021 by Fordham Intellectual Property, Media & Entertainment Law Journal; Katrina Geddes OCULARCENTRISM AND DEEPFAKES: SHOULD SEEING BE BELIEVING? The pernicious effects of misinformation were starkly exposed on January 6, 2021, when a violent mob of protestors stormed the nation's capital, fueled by false claims of election fraud. As policymakers wrestle with various proposals to curb misinformation online, this Article highlights one of the root causes of our vulnerability to misinformation, specifically, the epistemological prioritization of sight above all other senses (“ocularcentrism”). The increasing ubiquity of so-called “deepfakes”--hyperrealistic, digitally altered videos of events that never occurred--has further exposed the vulnerabilities of an ocularcentric society, in which technology-mediated sight is synonymous with knowledge. This Article traces the evolution of visual manipulation technologies that have exploited ocularcentrism and evaluates different means of addressing the issues raised by deepfakes, including the use of copyright law. INTRODUCTION 1043 I. Are Deepfakes Protected by Fair Use? 1046 A. Background 1046 B. Kim Kardashian Deepfake 1049 C. Jay-Z/Billy Joel Deepfake 1057 D. Democratizing Creative Production 1060 II. Individual and Collective Harms 1061 A. Our Vulnerability to Misinformation 1061 B. The History of Ocularcentrism 1064 III. Proposed Solutions 1074 CONCLUSION 1083 *1043 INTRODUCTION Since its unholy beginnings in pornography, deepfake technology has, understandably, been the subject of widespread criticism and outrage. Broadly speaking, a “deepfake” is a hyperrealistic video that has been digitally altered to depict an event or events that never occurred.1 At the individual level, pornographic and other harmful kinds of deepfakes can cause significant psychological and reputational harm.2 At the collective level, the dissemination of deepfakes affects our ability to differentiate authentic from inauthentic content, rendering us more vulnerable to misinformation.3 This effect, however, is not limited to deepfakes; photographs and videos have long been vulnerable to manipulation. The problem, then, is not deepfakes per se, but our uncritical and disproportionate reliance on technology-mediated sight, and our insistence that seeing is believing. The initial purpose of this Article is to understand the historical persistence of “ocularcentrism,” or the epistemological prioritization of sight above other human senses,4 and, secondly, to situate deepfakes within this social history--do deepfakes represent the limit of our tolerance for visual manipulation, and if so, why? Do they truly threaten visual truth in a way that earlier *1044 technologies have not? If so, should we abandon ocularcentrism--or cling to the credibility of visual evidence? To date, existing scholarship on deepfakes has failed to differentiate between, and tailor solutions for, the individual and collective harms associated with their dissemination. Such tailoring is needed to preserve the substantial utility that deepfakes © 2021 Thomson Reuters. No claim to original U.S. Government Works. 1 OCULARCENTRISM AND DEEPFAKES: SHOULD SEEING..., 31 Fordham Intell.... offer. Deepfake audio recreated the speech that John F. Kennedy intended to deliver shortly before his assassination, using recordings of 831 speeches he delivered in his lifetime, and offering hope to patients who have lost their voices to illness.5 Researchers have used deepfake technology to create animated, photorealistic avatars of deceased persons and portrait subjects.6 Museum visitors can interact with life-size deepfakes of long-dead artists, constructed from archival footage.7 Deepfake technology can be used to anonymize vulnerable sources,8 generate multilingual voice petitions,9 produce synthetic MRI images that protect patient privacy,10 synthesize news *1045 reports,11 improve video-game graphics,12 reverse the aging process,13 re-animate old photos,14 and elevate fanfiction.15 If, like most forms of technology, deepfakes are capable of both beneficial and harmful use, how should the technology be regulated to maximize its utility and minimize its harm? *1046 This Article will explore this question through the lens of copyright law and policy. The creation of deepfakes depends heavily on access to, and manipulation of, audiovisual content--much of which is protected by copyright law. Accordingly, copyright represents a natural lens through which to evaluate the unique social issues raised by the creation and dissemination of deepfakes. Part I will explain the technical process by which deepfakes are created and evaluate whether a deepfake video would constitute transformative fair use.16 Part II will discuss both the individual and collective harms generated by the dissemination of deepfakes, including the erosion of our ability to differentiate authentic from inauthentic content. It will interrogate the historical basis for the normative claim that seeing is believing and problematize the role of ocularcentrism in promoting both surveillance and misinformation. Part III will evaluate a variety of legal and regulatory measures that have been proposed to address the harms caused by deepfakes. The Conclusion will summarize the discussion contained within the Article and provide final thoughts. I. Are Deepfakes Protected by Fair Use? For now, this question remains theoretical; no judicial proceeding has yet determined whether fair use protects the creators of deepfakes from copyright infringement liability. So, the question becomes conditional: should deepfakes be protected by fair use? Would such protection be consistent with the evolution of fair use jurisprudence and the overarching policy objectives of the copyright regime? These are the questions that will be explored in this section. A. Background First, it is important to understand the technical process by which deepfakes are created. The term deepfake--a combination of “deep learning” and “fake”--generally refers to synthetic content *1047 created by an artificial neural network,17 but the term has colloquially been used to describe a broad spectrum of hyperrealistic content.18 At the sophisticated end of the spectrum, a recurrent neural network (“RNN”) can generate synthetic video footage of an individual from an audio recording.19 The process of mapping from a one-dimensional (audio) signal to a three-dimensional time-varying image is technically challenging, but bears substantial utility.20 For example, an individual who is hearing-impaired could lip-read a synthetic video generated from over-the-phone audio.21 Researchers from the University of Washington trained a RNN on seventeen hours of video footage of President Obama delivering 300 weekly addresses.22 From this corpus of video footage, they extracted 1.9 million video frames.23 For every output video frame, the RNN detects mouth landmarks (18 points along the outer and inner contours of the lip) to generate a sparse mouth shape.24 The mouth shape and lower region of the face are given texture before the synthesized mouth region is blended into the target video.25 The target video is then re-timed to ensure that the natural head motion matches the input audio.26 Essentially, the RNN maps mouth shapes from raw audio to create synthetic footage that can be composited into the mouth region of a target video for photorealistic results.27 © 2021 Thomson Reuters. No claim to original U.S. Government Works. 2 OCULARCENTRISM AND DEEPFAKES: SHOULD SEEING..., 31 Fordham Intell.... Another sophisticated technique for generating deepfakes is a generative adversarial network (“GAN”), which pairs a discriminative algorithm (which predicts a label, given certain features) and a *1048 generative algorithm (which predicts features, given a certain label).28 For example, a discriminative algorithm would try to predict whether a particular email should be classified as “spam” given its contents, whereas a generative algorithm would try to predict the features of an email that had already been classified as spam. Deepfakes are created by the interaction of these algorithms: the “generator” generates artificial images that resemble the images in the training set, and the “discriminator” evaluates these images for authenticity-- whether they came from the training set or not.29 As these algorithms interact, the generator learns to create sufficiently realistic images to fool the discriminator.30 A similar deep learning technique, known as Video Dialogue Replacement (“VDR”), was used to create a deepfake of Mark Zuckerberg discussing the profitability of personal data.31 Artists Barnaby Francis and Daniel Howe created the deepfake using the proprietary algorithm of an Israeli technology start-up known as Canny AI.32 Canny engineers trained their deep learning algorithm on a twenty-one second clip from the target video as well as video footage of the voice actor speaking, then reconstructed the frames in the target video to match the facial movements of the voice actor.33 No audio recordings of Zuckerberg were used.34 At the other end of the deepfake spectrum, less sophisticated actors can create “cheap fakes,” or lower-quality deepfakes, using *1049 consumer-grade software or simple video-editing techniques.35 For example,